US AI labs face new China-linked pressures

Alibaba (9988.HK) has ordered its employees to stop using Anthropic’s Claude Code for work and to uninstall Anthropic models from internal systems. Internal guidance set July 10, 2026 as the deadline for staff to comply, effectively cutting off one of the prominent corporate user bases for Claude Code inside China-linked operations.

The decision highlights growing operational friction between major Chinese technology firms and U.S. frontier AI providers. It also underscores how corporate compliance policies can quickly reshape access to leading AI tools when security or regulatory questions emerge.

Before Alibaba’s (9988.HK) move, a security researcher identified hidden detection and tracking logic in Claude Code. The code checked system timezone and proxy or network signals in an effort to identify users who might be located in China or connected to Chinese AI laboratories.

An Anthropic engineer, Thariq Shihipar, said this logic was an experiment introduced in March to prevent model distillation. He stated that the experiment had been removed or was scheduled for removal around July 1, 2026, following the scrutiny of the detection mechanism.

The existence and subsequent removal of this obfuscated logic has fueled mistrust around how U.S. AI models monitor and respond to usage that could be tied to rival AI development efforts.

In parallel with the code experiment, Anthropic has reported extensive attempts to extract its models’ capabilities through organized usage. The company alleged that roughly 25,000 fraudulent accounts generated more than 28.8 million exchanges with Claude between April 22 and June 5, 2026.

These exchanges are described as part of model “distillation” activity, in which repeated queries are used to replicate behaviors of a proprietary AI system. The scale of the reported activity illustrates the commercial and security stakes around safeguarding advanced model outputs.

Chinese authorities are reported to be discussing measures to limit overseas access to the country’s most advanced AI models. Proposals under consideration include tighter restrictions on foreign use, tougher penalties for technology leaks, and closer scrutiny of foreign investment, framed as national-security priorities.

Such steps would add a regulatory layer to the existing technical and commercial tensions, potentially constraining cross‑border AI collaboration and complicating compliance for global companies that rely on both U.S. and Chinese systems.

Alongside policy moves, Chinese open-weight models are gaining traction on performance and price. Beijing-based Z.ai’s GLM-5.2 can be downloaded and run locally and has been judged competent by security researchers on coding and vulnerability‑finding tasks.

Usage data show that after launch, GLM-5.2’s daily token volume grew about 27 times and its customer count increased about 80 times. Practitioners note that many Chinese open-source models are materially cheaper than leading U.S. proprietary systems, with estimates ranging from 60% to 90% lower cost for some workloads.

This pricing gap is prompting some U.S. users to route lower‑priority or less demanding tasks to cheaper Chinese models, while reserving premium frontier models for only the most complex jobs. The shift intensifies competitive pressure on U.S. labs and raises new questions about data governance across jurisdictions.